1
00:00:04,500 --> 00:00:08,490
So why is the game of
Jeopardy hard for a computer?

2
00:00:08,490 --> 00:00:13,090
We said earlier that Watson had
3,000 processors and a database

3
00:00:13,090 --> 00:00:16,700
of 200 million pages
of information.

4
00:00:16,700 --> 00:00:20,620
So shouldn't it be easy for
Watson to play Jeopardy?

5
00:00:20,620 --> 00:00:23,440
Unfortunately, Jeopardy
has a wide variety

6
00:00:23,440 --> 00:00:27,100
of categories which are
purposely made cryptic.

7
00:00:27,100 --> 00:00:30,960
While computers can easily
answer precise questions,

8
00:00:30,960 --> 00:00:34,290
like computing the square
root of a complicated number,

9
00:00:34,290 --> 00:00:38,470
understanding natural language
is hard for computers.

10
00:00:38,470 --> 00:00:42,570
As an example, suppose we ask
Watson to answer the question:

11
00:00:42,570 --> 00:00:45,620
Where was Albert Einstein born?

12
00:00:45,620 --> 00:00:47,770
Stored in its
database, Watson might

13
00:00:47,770 --> 00:00:51,610
have the following information:
"One day, from his city

14
00:00:51,610 --> 00:00:54,420
views of Ulm, Otto
chose a watercolor

15
00:00:54,420 --> 00:00:59,140
to send to Albert Einstein as a
remembrance of his birthplace."

16
00:00:59,140 --> 00:01:01,460
As a human, we can
probably figure out

17
00:01:01,460 --> 00:01:04,530
that the name of
the city is Ulm,

18
00:01:04,530 --> 00:01:07,360
but this is a hard
sentence to parse.

19
00:01:07,360 --> 00:01:10,030
How would you tell a computer
that Albert Einstein was

20
00:01:10,030 --> 00:01:12,200
born in Ulm using
just this sentence?

21
00:01:14,820 --> 00:01:15,320
OK.

22
00:01:15,320 --> 00:01:17,490
So how about if we
just store answers

23
00:01:17,490 --> 00:01:21,330
to all possible questions that
could be asked on Jeopardy?

24
00:01:21,330 --> 00:01:24,140
Unfortunately, this
would be impossible.

25
00:01:24,140 --> 00:01:28,600
An analysis of 200,000
previous Jeopardy questions

26
00:01:28,600 --> 00:01:32,070
yielded over 2,500
different categories,

27
00:01:32,070 --> 00:01:36,270
and new questions are created
on Jeopardy all the time.

28
00:01:36,270 --> 00:01:38,570
Well, OK, then let's
just search Google

29
00:01:38,570 --> 00:01:40,890
for the answer to the question.

30
00:01:40,890 --> 00:01:43,800
Unfortunately, no links
to the outside world

31
00:01:43,800 --> 00:01:46,240
are permitted on
Jeopardy, and this rule

32
00:01:46,240 --> 00:01:48,610
applied to Watson as well.

33
00:01:48,610 --> 00:01:51,700
And even if Watson could search
the internet for the answer

34
00:01:51,700 --> 00:01:54,940
to a question, it can
take considerable skill

35
00:01:54,940 --> 00:01:58,950
to find the right web page
with the right information.

36
00:01:58,950 --> 00:02:01,380
So instead, Watson
used analytics

37
00:02:01,380 --> 00:02:03,780
to answer the
Jeopardy questions.

38
00:02:03,780 --> 00:02:07,180
Watson received each
question in text form.

39
00:02:07,180 --> 00:02:10,250
Normally, the players see
and hear the questions,

40
00:02:10,250 --> 00:02:11,860
but Watson couldn't
hear anything,

41
00:02:11,860 --> 00:02:16,260
so they decided to feed him
the questions in text instead.

42
00:02:16,260 --> 00:02:18,990
With the question
in text form, IBM

43
00:02:18,990 --> 00:02:22,660
was able to use text analytics
and other analytical methods

44
00:02:22,660 --> 00:02:25,600
to make Watson a
competitive player.

45
00:02:25,600 --> 00:02:28,200
Overall, they used 100
different techniques

46
00:02:28,200 --> 00:02:31,760
for analyzing natural
language, finding hypotheses

47
00:02:31,760 --> 00:02:34,860
for the questions, and
ranking these hypotheses

48
00:02:34,860 --> 00:02:37,829
to pick an answer.

49
00:02:37,829 --> 00:02:41,910
Watson had a huge database of
sources and several basic tools

50
00:02:41,910 --> 00:02:44,310
to help understand language.

51
00:02:44,310 --> 00:02:48,270
The database consisted of a
massive number of data sources,

52
00:02:48,270 --> 00:02:52,970
like encyclopedias,
texts, manuals, magazines,

53
00:02:52,970 --> 00:02:56,700
and downloaded
pages of Wikipedia.

54
00:02:56,700 --> 00:03:00,220
One of the tools Watson
had was a lexicon,

55
00:03:00,220 --> 00:03:03,880
which describes the relationship
between different words.

56
00:03:03,880 --> 00:03:07,250
For example, the lexicon
could tell Watson

57
00:03:07,250 --> 00:03:10,560
that water is a clear liquid,
but not all clear liquids

58
00:03:10,560 --> 00:03:12,220
are water.

59
00:03:12,220 --> 00:03:14,500
Another tool Watson
had was a part

60
00:03:14,500 --> 00:03:17,490
of speech tagger and parser.

61
00:03:17,490 --> 00:03:21,400
This would identify
functions of words in text.

62
00:03:21,400 --> 00:03:24,190
For example, it would
know that the word "race"

63
00:03:24,190 --> 00:03:27,329
can be used as a verb or a noun.

64
00:03:27,329 --> 00:03:30,030
"The students had to
race to catch the bus"

65
00:03:30,030 --> 00:03:34,210
uses race as a verb, while
"Please indicate your race"

66
00:03:34,210 --> 00:03:37,400
uses race as a noun.

67
00:03:37,400 --> 00:03:40,400
Then, using these data
sources and tools,

68
00:03:40,400 --> 00:03:42,400
Watson would answer
a question by going

69
00:03:42,400 --> 00:03:44,930
through four major steps.

70
00:03:44,930 --> 00:03:47,730
The first step is
question analysis,

71
00:03:47,730 --> 00:03:49,640
where Watson tries
to figure out what

72
00:03:49,640 --> 00:03:51,990
the question is looking for.

73
00:03:51,990 --> 00:03:55,350
The second step is
hypothesis generation,

74
00:03:55,350 --> 00:03:57,380
where Watson searches
the information

75
00:03:57,380 --> 00:04:00,410
sources for possible answers.

76
00:04:00,410 --> 00:04:02,740
After this, Watson
moves on to step

77
00:04:02,740 --> 00:04:06,330
three, when the different
hypotheses are scored.

78
00:04:06,330 --> 00:04:08,100
This means that a
confidence level

79
00:04:08,100 --> 00:04:11,570
has to be computed
for each answer.

80
00:04:11,570 --> 00:04:14,490
The final step is
ranking the hypotheses

81
00:04:14,490 --> 00:04:17,890
to look for a
highly-supported answer.

82
00:04:17,890 --> 00:04:21,320
In the next two videos, we'll go
through how each of these steps

83
00:04:21,320 --> 00:04:22,870
work.